TWI828416B - Machine learning in metrology measurements - Google Patents

Abstract

Metrology methods and targets are provided, that expand metrological procedures beyond current technologies into multi-layered targets, quasi-periodic targets and device-like targets, without having to introduce offsets along the critical direction of the device design. Machine learning algorithm application to measurements and/or simulations of metrology measurements of metrology targets are disclosed for deriving metrology data such as overlays from multi-layered target and corresponding configurations of targets are provided to enable such measurements. Quasi-periodic targets which are based on device patterns are shown to improve the similarity between target and device designs. Offsets are introduced only in non-critical direction and/or sensitivity is calibrated to enable, together with the solutions for multi-layer measurements and quasi-periodic target measurements, direct device optical metrology measurements.

Description

Machine learning in metrology and measurement

The present invention relates to the field of metrology and, more particularly, to deriving metrological measurements and improving targets and simulation results.

The measurement objectives and methods are intended to derive measurements that represent the production accuracy of the designed device. Metrology faces the challenge of producing measurable signals that reflect the accurate properties of the device at a rate high enough and at an effective area low enough to minimize production barriers. Current metric overlay (OVL) algorithms use special targets with periodic structures in both layers that are offset differently in different target cells.

U.S. Patent Application Publication No. 2016/0266505, which is incorporated herein by reference in its entirety, discloses on-device optical measurement and real-time model-based measurement algorithms.

U.S. Patent Application Publication No. 2009/0244538, which is incorporated herein by reference in its entirety, discloses a lithography apparatus configured to transfer a pattern from a patterning device to a substrate, wherein a reference grating set Disposed in the substrate, the reference set includes two reference gratings with line elements in a first direction and one reference grating with line elements in a second vertical direction. US 2009/0244538 requires the same (or very similar) design for x and y in order to calculate the sensitivity in one direction and apply it in the second direction. However, this x-y design symmetry is broken in the device due to electrical functionality requirements (and also because the lithography process in the critical layer is asymmetric).

U.S. Patent Application Publication No. 2011/0255066, which is incorporated by reference in its entirety, discloses the use of multiple fields assuming constant overlay sensitivity of targets across a wafer field, ignoring intra-field process variations. Use multiple targets to measure overlap.

Young-Nam Kim et al. 2009 (Device-based in-wafer critical dimension and overlay metrology, Optics Letters 17:23, 21336-21343), which is incorporated by reference in its entirety, discloses a model-based in-wafer optical metrology. technology that allows direct measurement of both critical dimension and overlay displacement errors during the DRAM fabrication process performed on actual semiconductor devices without the need for special target structures.

The following provides a simplified summary for an initial understanding of the invention. This summary does not necessarily identify key elements nor limit the scope of the invention, but is merely an introduction to the description that follows.

One aspect of the invention provides a method of measuring a metrology parameter directly on a device, the method comprising: (i) measuring at least one metrology parameter from at least a portion of a device design, the at least portion being selected to be along the having a plurality of irregular repeating units in at least one direction of the portion having specified characteristics such as different sets of lines and cuts, and (ii) applying at least one machine learning algorithm to use at least one of: one to calibrate sensitivity to: a diffraction order intensity normal to the at least one direction; an introduced offset along a non-critical direction; a target cell adjacent to the device portion having an introduced offset; and At least one sensitivity calibration target, wherein the measurements are performed as scatterometry on a plurality of targets to provide layer-specific metrology parameters, at least one of the targets being a portion of the device having N>2 overlapping layers A portion, wherein the plurality of targets includes at least one of the following: N pairs of cells, each pair having opposite offsets at a different layer; N cells having a selected predetermined offset; N cells or less having selected offsets configured to utilize optical beam information; and N cell calibration targets located next to a number of overlay targets between N-1 and two .

One aspect of the present invention provides a method, which includes: configuring a multi-layer metrology target to have a plurality of M target cells located on at least three target layers, N≤M, each cell in each layer There is at least one periodic structure in each cell, and the periodic structures of each cell are configured to be offset relative to each other by a specified offset; at least M differential measurements are measured from the multi-layer metrology target in a scattering measurement manner. signals; and applying at least one machine learning algorithm to the differential signals and the specified offsets by matching scatter measurement overlay (SCOL) metrology parameters to the differential signals and the specified offsets. The SCOL measurement parameters are calculated from the M measurements of the multi-layered measurement target by moving a related set of M equations to solve.

One aspect of the present invention provides a method that includes calibrating at least one sensitivity parameter by applying at least one machine learning algorithm using a shift in at least one of the following without moving along a Measuring at least one measurement parameter in at least one target cell with the critical measurement direction introducing a predetermined offset into the at least one target cell: (i) an orthogonal non-critical measurement direction; and (ii) At least one additional target cell in addition to the at least one target cell.

One aspect of the invention provides a method of measuring metrological parameters directly on a device, which synergistically combines the above-mentioned methods.

These, additional and/or other aspects and/or advantages of the invention are set forth in the detailed description that follows; may be inferred from the detailed description; and/or may be learned by practice of the invention.

Cross-references to related applications

This application claims the rights and interests of U.S. Provisional Patent Application No. 62/546,509, filed on August 16, 2017, which is incorporated herein by reference in its entirety.

Before setting out the detailed description, it may be helpful to state definitions of specific terms that will be used below.

The term "metrology target" or "target" as used herein in this application is defined as any structure designed or produced on a wafer that is used for metrology purposes. Specifically, the overlay target is designed to achieve measurement of the overlay between two or more layers produced in a film stack on a wafer. Exemplary overlay targets are scatter measurement targets measured by scatter measurements at the beam plane and/or field plane, and imaging targets.

Exemplary scatterometry targets may include two or more periodic or non-periodic structures (referred to in a non-limiting manner as gratings) located at different layers and may be designed and produced on top of each other (referred to as "Grating on grating") or adjacent to each other from a vertical perspective (called "side-by-side"). The target design contains one or more cells, the term "cell" as used herein in this application being defined as a portion of a target used to derive a measurement signal. Common scatter measurement targets are called SCOL (Scatter Measurement Overlay) targets, DBO (Diffraction Based Overlay) targets, etc. Common imaging targets are called Box-in-Box (BiB) targets, AIM (Advanced Imaging Metrology) targets, AIMid targets, Blossom targets, etc. Note that in the present invention, SCOL is related to model-free in the sense that the details of the measured stack must be unknown before measurement and the target does not necessarily need to be modeled in order to extract parameters. Note further that the present invention is not limited to any one of these specific types, but may be implemented with respect to any target design.

Target elements include periodic structures, such as gratings, with elements repeating at one or more pitches. Certain metrology targets exhibit an "induced shift," also known as a "designed shift,""designedshift," or "designed misalignment," which as used herein in this application is an intentional shift. Or the overlap between the periodic structures of the target. The term "overlap" as used herein in this application is defined as the overall offset, that is, the intended offset between two layers of an object or device plus an unintentional offset. Therefore, unintentional offsets can be derived by subtracting a given shift from the measured alignment. Overlay targets are typically designed to have pairs of cells, with the cells of each pair having equal and opposite predetermined offsets, denoted ±f ₀ .

The term "periodic" as used herein in relation to an object or an object structure is defined as having a cyclic pattern. The term "strictly periodic" as used herein in this application refers to an object or an object structure having a recurring unit cell that is the same at all cycles. The term "quasi-periodic" as used herein in this application refers to an object or an object structure having a recurring pattern that does not have a recurring unit cell but exhibits underlying A basic pattern of elements, such as a grid, and aberrations (eg, changes in the length, width, or detail of the cyclic pattern and/or changes in grid parameters and characteristics). These aberrations may be carefully chosen (as explained) and/or may be random and/or may reflect design considerations. The effect of aberration on the signal (i.e., the difference between the signal derived from the self-quasi-periodic target and the equivalent strictly periodic target) is referred to in this application as "noise", which may have random and system components. Noise may be considered part of the inaccuracy defined below (for strictly periodic designs) and/or may be treated independently.

The term "device" or "device design" as used herein in this application is defined as any portion of a wafer that provides an operating electronic circuit, such as a memory device or a logic device. As used herein in this application, the term "critical direction" is defined as a direction in a device design that is sensitive to small shifts (eg, on the order of 1 nm) between layers, where if such shifts occurs, the device may malfunction. As used herein in this application, the term "non-critical direction" is defined as a direction in which small shifts (eg, on the order of 1 nm) can be tolerated in the device design, and if such shifts occur, then The device does not malfunction.

The term "measurement direction" as used herein in this application is defined as one of the directions along which the measurement overlaps. When using periodic targets, there must be periodicity in the measurement direction. The spacing is on the order of up to several thousand nanometers. In addition to this coarse spacing, there may also be a generally much smaller segmental spacing of the features and/or a coarse spacing in the orthogonal direction.

Note that although the invention is directed to optical illuminating radiation, the invention extends to applications where the illuminating radiation is at very short wavelengths (such as x-rays) and/or includes particle beams.

With respect to metrological measurements, the term "differential signal" as used herein in this application is defined as the intensity between two signals measured from a target, such as the +1 and -1 diffraction order signals. Difference. The term "sensitivity" as used herein in this application is defined as the overall offset or overlap between a differential signal and a periodic structure along a respective measurement direction. Ratio or relationship. The term "inaccuracy" as used herein in this application is defined as the difference between a result of a measurement and the true value of the measured measurement (the measurand). It should be emphasized that although the models presented are mostly linear for reasons of clarity, linearity is non-limiting in the sense that the algorithm can be adapted to exploit non-linear models and is therefore also part of the invention.

In the following description, various aspects of the invention are set forth. For purposes of explanation, specific configurations and details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the present invention may be practiced without the specific details presented herein. Additionally, well-known features may have been omitted or simplified so as not to obscure the invention. With specific reference to the drawings, it is emphasized that the details shown are by way of example and for the purpose of illustrative discussion of the invention only and are intended to provide what is believed to be most useful and to enable the principles and conceptual aspects of the invention to be understood. The instructions are presented in an easy-to-understand manner. In this regard, the structural details of the invention are shown only to the extent necessary to obtain a basic understanding of the invention, and the illustration by means of the drawings enables those skilled in the art to understand how several aspects of the invention may be embodied in practice. form.

Before at least one embodiment of the present invention is explained in detail, it is to be understood that this invention is not limited in its application to the details of construction and the arrangement of components set forth in the following description or illustrated in the drawings. The invention is capable of other embodiments or of being practiced or carried out in various ways. Furthermore, it is to be understood that the phraseology and terminology employed herein are for the purpose of description and should not be regarded as limiting.

Unless otherwise specifically stated, as is apparent from the above discussion, it should be understood that throughout this specification, discussions utilizing terms such as "processing," "computing, calculating," "determination," "enhancement," or the like means a computer or computing system or similar electronic computing device that manipulates data represented as physical (such as electronic) quantities in the computer system's registers and/or memory and/or transforms such data into similar representations The actions and/or processes that are physical quantities in the memory or register of a computer system or other such information storage, transmission or display device).

Embodiments of the present invention provide efficient and economical methods and objectives for direct use of device design, specifically the use of model-free far-field optical metrology to measure metrology parameters, specifically overlays. Specifically, the following disclosure overcomes three major obstacles that have prevented prior art direct device measurement, namely, the multi-layered nature of the device design, the aperiodic nature of the device design, and the need to avoid introducing intended offsets into the actual device design. (so as not to damage its electrical properties and performance) inherent constraints.

Providing metrology methods and targets that extend metrology procedures beyond current technology into multi-layer targets, quasi-periodic targets, and device-like targets without introducing offsets along critical directions of device design. Discloses the application of machine learning algorithms to measurement and/or measurement simulation of measurement targets to derive measurement data such as overlays from multi-layered targets, and provides corresponding configurations of targets to achieve such measurements. Quasi-periodic targets based on device patterns were shown to improve the similarity between targets and device designs. Introducing offset and/or calibration sensitivity only in non-critical directions enables direct device optical metrology measurements along with solutions for multi-layer metrology and quasi-periodic target metrology.

These methods and objectives are exemplified for Scattering Optical Overlay (SCOL) measurements of device structures, which is a rapid non-destructive overlay (OVL) technique. The main limitation of this fast non-destructive OVL technique is the need for special targets due to its limited resolution. These measurement targets may have poor correlation with the actual device structure due to large deviations in its design and location. Therefore, it is desirable to directly measure the device to better reflect its OVL and other parameters that may be of interest. The section labeled "Multiple Layer Targets" provides methods to achieve SCOL measurements of multiple overlapping parallel patterns. The section labeled "Quasiperiodic Targets" explains how to deal with device designs that lack a unit cell and are not strictly periodic and how to deal with the signal-to-noise ratio in SCOL. Finally, the section labeled "Avoiding offsets in device targets" presents an innovative approach to sensitivity calculations without damaging electrical properties. The disclosed methods, algorithms, and target designs are collaboratively combined into a complete solution for on-device optical OVL measurement.

The inventors have discovered that although U.S. Patent Application Publication No. 20160266505 discloses solutions to problems that are problematic in the context of overlay measurements related to device characteristics (e.g., having complex multi-patterned structures, not being periodic, having Analytical methods and models (mostly physics-based) of the electrical properties of the given offset damage required for SCOL targets and provide corresponding target designs - but a complementary or alternative method for self-measurement signal extraction information could be to use machines learning methods. For example, a machine learning algorithm can be trained on a cluster of signals and corresponding desired measurements to find correlations and functions used to extract the required parameters. For example, machine learning algorithms can be used to reduce the number of cells required in a multi-level target design, possibly to a single cell.

In certain embodiments, machine learning algorithms can be used to derive overlay sensitivity calibrations based on real-time information, which can come from different sources, such as: additional diffraction orders, orthogonal diffraction orders, a first order with different designs. Two goals, etc. Machine learning methods can be applied with respect to other methods for extracting separate information to provide mutual reinforcement of the efficiency of the respective methods. multi-layered goals

First, multi-layer targets are revealed together with corresponding measurement and signal derivation algorithms and machine learning algorithms based on them, so that the measurement of these multi-layer targets has no or smaller processing capacity/effective area penalty. The objectives are discussed in a non-limiting manner in a one-dimensional context and for the scatterometric overlay (SCOL) objective. These goals and methods are expected to improve current technology in at least the following ways: (i) Enable the design and measurement of optical alignment (OVL) of in-die targets that better reflect device alignment ; (ii) provides greater flexibility in the design of target dummification (for example) by allowing features to be parallel to the measurement direction; and (iii) can improve effective area and/or throughput specifications.

The present invention overcomes the limitations of standard scatterometry overlay algorithms, which require that the overlay between two gratings is the only source of symmetry violation. When additional gratings are present, their relative offsets can cause the signal to change in ways that cannot be processed using standard overlay algorithms. This contaminates the original two-layer stack signal and results in an inaccurate measurement. In addition, machine learning algorithms achieve reductions in the number of cells per target, extraction of more information and/or faster extraction of information from self-measurement, and improved target design regarding the parameters of the machine learning algorithms. Machine learning algorithms can be trained on target designs based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Figure 1A is a high-level schematic overview illustration of a multi-layered target and a method of measuring the same, in accordance with certain embodiments of the present invention. Prior art overlay targets and algorithms 90 relate to two-tier targets and respective overlay algorithms using opposite predefined offsets ± along each measurement direction. f ₀ has at least two measurement cells. It should be emphasized that the application of prior art algorithms to targets with more than two layers creates an excessive number of variables due to the interaction between the illumination layer and the target layer. For example, prior art algorithms provide two equations (corresponding to two target cells with opposite offsets) to derive two variables (relative displacement between layers (i.e., overlay)), and make the overlay For the coefficient A ) related to the differential signal. In the case of N > 2 target layers, the two equations provided by the prior art overlay algorithm are insufficient to derive the overlay between N layers.

The present invention proposes a method 100 and an overlay target 290 having three or more layers, which achieves the extraction of various metrological parameters (herein represented by overlays) from multi-layer targets.

Particular embodiments propose using a modified overlay algorithm as a method 100A to measure a target 201 with a plurality of - N 2-cell targets, where each of the N layers has a cell pair, all cells are the same, but for each layer there is a corresponding cell pair with a given offset ±f ₀ along the measurement direction. For example, each target may include two periodic layers configured to measure overlap between respective layers. Overlay may be derived from measurements of target 201 by a modified version of the prior art algorithm 90 that considers intermediate layers. However, this innovative solution has the following disadvantages: (i) the targets are very different from the device (increased device and metering offsets) and (ii) the targets are less printable, even on the scribed line, since they are relatively Stay away from the window to be treated. Although it is possible to measure all overlay systems using the target 201 by the modified SCOL algorithm 100A, the following methods (100B to 100D) prove to be significantly more efficient in terms of saving effective area and reducing MAM time.

Machine learning algorithms 150 may be applied in association with any of methods 100A-100D to enhance derivation accuracy, reduce errors, speed up overall measurement time, reduce the number of cells required 155 and/or achieve device-to-class goals Model-free real-time measurement. In certain embodiments, a machine learning algorithm may be trained on a target design based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Certain embodiments use half the number of cells, that is, N cells (or more) for measuring N layers by judiciously selecting offsets based on parameters of a machine learning algorithm applied to simulated measurements Target. Note that the disclosed methods can be used to replace or enhance methods derived from the analytical methods set forth in US Patent Application Publication No. 20160266505. The inventor has found that the analysis method ensures the applicability and good convergence characteristics of the machine learning algorithm, and at the same time, the machine learning algorithm can avoid some of the assumptions involved in the analysis method.

Certain embodiments use even fewer cells, i.e., N-1 cells (or more) and possibly even less than N-1 cells for utilizing optical information by judicious selection of offsets. (Information in the signal measured at the optical plane relative to the target of the optical system of the metrology tool) to measure the N- layer target. Note that using less than N cells per target is advantageous in terms of saving effective wafer area and reducing MAM (move-acquire-measure) time. Note that the disclosed methods can be used to replace or enhance methods derived from the analytical methods set forth in US Patent Application Publication No. 20160266505. The inventor has found that the analysis method ensures the applicability and good convergence properties of the machine learning algorithm, and at the same time, the machine learning algorithm can avoid some of the assumptions involved in the analysis method. Particular embodiments of the present invention include objects with n<2N cells, which are achieved by applying machine learning algorithms to extract overlay information from n<2N cells. In certain embodiments, the targets include a single cell per target, and the targets are achieved by applying machine learning algorithms to achieve model-free real-time optical overlay measurements of the single cells.

Certain embodiments use even fewer cells, i.e., as few as two cells (or more) 300 located near the device to exploit the optical beam information by judicious selection of offsets and by using Additional calibration target 200, method 100D, targets 200, 300, and Figure 1B below positioned at an area further away from the device (e.g., on the reticle) (note that method 100D uses and further develops methods 100B, 100C and targets 200, 300 ) to measure N- layer targets—as explained and illustrated in U.S. Patent Application Publication No. 20160266505. Note that the disclosed methods can be used to replace or enhance methods derived from the analytical methods set forth in US Patent Application Publication No. 20160266505. The inventor has found that the analysis method ensures the applicability and good convergence characteristics of the machine learning algorithm, and at the same time, the machine learning algorithm can avoid some of the assumptions involved in the analysis method.

In a specific embodiment, the targets include a single cell per target, and model-free real-time optical overlay measurements of the single cells are achieved by applying machine learning algorithms.

Target 290 may be measured under different and possibly multiple hardware and illumination configurations, for example, using different wavelengths and/or illumination modes, using different polarizations, using different apodizers or different other components in the optical system, 201, 200, and 300 to enhance calibration and measurement, especially when applying methods 100C and 100D. Machine learning algorithm 150 may be applied in combination with or in place of any of methods 100A-100D to potentially further reduce the number of cells 155 required.

Figure 1B is a high-level schematic illustration of two types of multi-layer targets 200, 300 and their measurement methods 100 in accordance with certain embodiments of the present invention. The machine learning algorithm 150 may enhance or replace the first method 100B (which uses an analysis of differential signals from multiple grating objects) and/or the second method 100C (which uses an approximate decomposition of overlays reported by standard algorithms ). Both methods rely on the use of plenoptic information to extract additional required information. For each method, a non-limiting example of a three-level objective is shown where computations are required to infer overlapping values. Note that three layers of targets are used in each measurement direction, that is, where N = 3, and for measuring X and Y in two directions, six cells are used to measure one of the three layers in both directions. Overlap. Obviously, the target can be designed similarly to provide measurements along only a single (critical) direction. Machine learning algorithm 150 may replace or enhance methods 100B, 100C, which methods are described in detail in US Patent Application Publication No. 20160266505. Machine learning algorithms 150 can further be used to reduce the number of cells 155 required.

Methods 100B, 100C are described below for measuring multi-layer SCOL metrology targets (ie, targets using a plurality of periodic structures (hereinafter related in a non-limiting manner to gratings) designed to be printed on wafer 60) Two non-exhaustive and non-limiting examples of method 100 of overlapping parameters. Possible combinations of methods 100B and 100C are suggested after explaining the principles of each method. Figure 2 is a high-level schematic illustration of a multi-layered object 200 in accordance with certain embodiments of the present invention. Figures 3A and 3B are high-level schematic illustrations of a multi-layered object 300 in accordance with certain embodiments of the invention.

Figure 2 schematically illustrates an object 200 that includes N cells 220 in N layers 210, each cell 220 having at least one periodic structure 230. Note that the selection of the same number (N) of cells and layers is only made to simplify the explanation below and does not limit the scope of the present invention. The number of cells can also be larger or smaller than the number of layers (see additional derivation below for the latter possibility). The periodic structures 230 are overlapping (on top of each other) due to predefined (given) offsets ( fi _,n , related to cell i and layer n ) between cell layers 210 and by the metrology method 100 Uncontrolled (non-predetermined) deviation of the target ( δ _i,n , related to cell i and layer n ) is represented. Both offsets are derived from signal 205 by estimating respective overlaps affected by the two types of offsets. The measured signal in the first-order scattering measurement stack is the differential signal D 205 , which is the intensity difference between +1 and -1 diffraction orders when illuminating a target cell 220 . Differential signal D 205 is used as a non-limiting example because the disclosed method can be used to measure other diffraction orders and the difference between the derived metrological measurements. The analysis method of measured signals provided in US Patent Application Publication No. 20160266505 can be replaced or supplemented by the disclosed machine learning algorithms, which can make it possible to avoid the assumptions involved in the analysis method. Certain assumptions and analysis methods suggest that may assist in the convergence of the disclosed machine learning algorithms. Machine learning algorithms 150 can further be used to reduce the number of cells 155 required.

A standard overlay target has two layers and therefore two unknown parameters, and uses two cells to provide the required two measured signals. Method 100 includes developing a new formalism required to handle more than two overlapping gratings in order to distinguish between the effects of offsets of different layers on the signal. The inventors note that this challenge has not yet been addressed due to the high-order complexity involved in the design of the target, theoretical analysis, and practical measurement procedures, all of which are disclosed in this disclosure. In the following sections, two innovative formalisms (corresponding to procedures 100B and 100C) are illustrated and demonstrated by simplified non-limiting models. The inventors note that one skilled in the art can readily extend such models to include additional contributions (eg, higher diffraction orders) that are therefore also considered part of the invention.

Method 100B uses inspection of differential signals 205 from target 200 to obtain overlap 235 (notation - OVL). Two variations of method 100B are presented—one variation assumes that previous values of OVL are known, and a more advanced variation uses optical information without requiring a priori knowledge of OVL in order to obtain all OVL in the design. value.

A second, more advanced variation of method 100 (referred to herein as method 100C) overcomes the problem of all previous OVL values by combining information from all pixels within the aperture and using the fact that the previous OVL value does not depend on the pixel position. need. The latter observation is used to define a cost function Ω that has a zero difference relative to any value of OVL (in a non-limiting manner, an example is relative to OVL ₁ ), as in US Patent Application Publication No. 20160266505 expressed and developed. Machine learning algorithms 150 may be applied in combination with or in lieu of methods 100B and/or 100C to enhance derivation accuracy, reduce errors, speed up overall measurement time, reduce the number of cells required 155 and/or achieve the goal of eliminating similar devices. Model real-time measurement measurement.

Figures 3A and 3B are high-level schematic illustrations of a multi-layered object 300 in accordance with certain embodiments of the invention. 3A and 3B illustrate a three-layer object 300 in a non-limiting manner. Those skilled in the art can implement the disclosed principles to multi-layer objects, and therefore these multi-layer objects are also considered to be part of the disclosed invention.

Multi-layer targets have not been used in prior art because the extra layer (more than two layers) is an additional source of symmetry breaking that contaminates the overlay signal from the two layers and produces inaccurate measurements. In the following, one or more additional layers are considered sources of inaccuracy and their effects on the signal are characterized. This representation is used to (i) eliminate the inaccuracy contribution of the additional layer to the overlay of an original two-layer object (which may be arbitrarily selected in object 300); and (ii) calculate the bias of the additional layer relative to the original layer. shift. These are part of method 100D (see Figure 1A), which may replace and/or enhance methods 100B and 100C of measuring multi-layer target 200 described above. In particular, the distinction between objects 200 and 300 is for illustrative purposes only and does not limit the scope of the invention, as it is clear that multi-layered objects may be designed to combine features of objects 200 and 300 .

Figure 3A schematically illustrates a non-limiting scenario in which the top layer 310 and the middle layer 320 are each considered to be the original layer for which a stack of pairs is to be calculated, while the bottom layer 330 (which may be replaced by multiple layers) is considered a source of inaccuracy. The effect of the offset of the bottom layer relative to the intermediate layer is similar to a symmetry break due to sidewall angular asymmetry. Figure 3B schematically illustrates a three-layer target 300 with the designation of the diffraction electric field as defined below after illumination I. For simplicity, it is assumed in a non-limiting manner that the periodic structures in layers 310, 320, 330 are parallel gratings with the same pitch. Assume further in a non-restrictive way that the order system before the diffraction electric field —The first-order signal is diffracted from the top grating 310, - the field transmitted through the top grating 310, exiting the middle grating 320 by first order diffraction and transmitted through the top grating 310 and - Fields transmitted through the top grating 310 and the middle grating 320 respectively, diffracted out of the bottom grating 330 by first order and transmitted through the middle grating 320 and the top grating 310 respectively, as illustrated in Figure 3B. The corresponding intensity of each of these fields is , and .

Under the assumptions stated above, the analysis method of measured signals provided in U.S. Patent Application Publication No. 20160266505 can be replaced or supplemented by the disclosed machine learning algorithms, which can enable Some of the assumptions involved in the analysis methods are avoided and the analysis methods suggest or may assist in the convergence of the disclosed machine learning algorithms. For example, the analysis method of US Patent Application Publication No. 20160266505 illustrates that the measured overlap at each pixel can be separated into an overlap between top layers and one that depends on the offset of the bottom layer. In another example, an analytical method separates variable momentum dependence from bottom layer offset. These instructions can support the application of machine learning algorithms and help identify useful parameters of machine learning algorithms that will be used as inputs to measurement and target design methods.

Calibration can be applied to a full sample or to subsequent wafers. Alternatively or in addition, differential signal analysis (methods 100B, 100C possibly enhanced by machine learning algorithms) may be applied to the calibration target. (iii) Subsamples can be measured across the wafer at dedicated targets positioned adjacent to external reference targets and the match between targets can be optimized. Calibration can be applied to a full sample or to subsequent wafers. (iv) A principal component analysis (PCA) can be performed on the sub-samples to derive relative measurements , and the absolute value can be calculated since OVL is obtained from the multi-cell target, as explained above. In certain embodiments, the machine learning algorithm 150 may be used with respect to PCA (eg, before PCA, after PCA), i.e., using principal components of PCA for the machine learning algorithm and/or possibly using a machine learning algorithm to improve the derivation of principal components.

Returning to Figure IB, certain embodiments of method 100D may combine the use of methods 100B and 100C. For example, two-cell targets 300 may be printed on a wafer along with a smaller number of (dedicated, calibrated) multi-cell targets 200 (ie, targets with three cells or more). The multi-cell target 200 may be sampled to be obtained by an overlay decomposition method or by examining the resulting differential signals and obtained from the differential signal analysis method 100B And get function. Machine learning algorithms 150 may be applied in connection with the method 100D described below and in U.S. Patent Application Publication No. 20160266505 to enhance derivation accuracy, reduce errors, speed up overall measurement time, and reduce the number of cells required 155 And/or achieve model-free real-time measurement of similar devices. In certain embodiments, a machine learning algorithm may be trained on a target design based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Note that the assumption of three periodic structures (gratings) is a non-limiting assumption presented here for simplicity only. In the case of two (or more) measurement directions, separate periodic structures can be added. To save effective wafer area, the calibration multi-cell targets 200 can be relatively few, while measurements of the two-cell targets 300 can be performed using calibrations derived from the two-cell targets.

More specifically, by studying the OVL distribution across the beam in a linear array of bottom gratings, it is possible to extract (For example, using principal component analysis and/or machine learning algorithms 150). For all other sites on the wafer, "OVL" can be separated into OVL ₂₃ ("common to all pixels") and OVL ₁₂ ("per-pixel inaccuracy"), as in U.S. Patent Application Publication No. 20160266505 elaborated, thereby enabling dedicated multi-layer target measurements using two cells (out of three layers). The respective metrological measurements of any of the targets 290, 201, 200, 300 and the application of the machine learning algorithm 150 to any of these targets are also considered part of this invention.

Figure 4 is a high-level flow diagram illustrating a method 100 in accordance with certain embodiments of the invention. Method 100 may be performed, at least in part, by at least one computer processor. Computer program products and corresponding metrology modules are provided that include a computer-readable storage medium having a computer-readable program embodied thereon and configured to at least partially perform the method 100 . It also provides target design files and target measurement.

Method 100 may include any of the following (either alone or in combination): modifying the current OVL algorithm to operate on N pairs of cells ( 2N cells), each pair having opposite biases in a layer. Shift (Method 100A) (e.g., each target has two periodic layers configured to measure overlap between respective layers); using only N cells designed to achieve Derivation of specific defined offsets for measured overlays (Method 100B); use of photon information to reduce the number of cells required to N-1 (Method 100C), and use of calibration targets to target specific overlays The number is reduced to less than N-1 , possibly as low as 2 targets per N levels (Method 100D).

The method 100 may further include enhancing and/or replacing any of the analysis methods (100A-100D) with a machine learning algorithm 150 (stage 102). In certain embodiments, method 100 may include applying a machine learning algorithm to calibrate overlay sensitivity and possibly reduce the number of cells to less than 2N (stage 105).

Method 100 may include configuring a multi-layer metrology target with a plurality (M) of target cells located on at least three target layers, N≤M, each cell having at least one periodic structure in each layer. (Stage 110) and the periodic structure of each cell is configured to be offset relative to each other by a specified offset (Stage 115). The method 100 may include measuring at least M differential signals in a scatterometry manner from a multi-layer metrology target (stage 120), and applying a machine learning algorithm to automatically derive the SCOL metrology parameters by relating them to the differential signals and specified offsets. SCOL metrology parameters are calculated for the M measurements of the multi-layer metrology target (stage 130). The SCOL metrology parameters can be the overlap between N layers. The applied machine learning algorithm 130 may be executed sequentially for successive layers to calculate SCOL metrology parameters (stage 132), for example, as in the first variation 100B of the method 100 set forth above. For example, the SCOL metrology parameters may be overlays between N layers, and machine learning algorithms may be applied with respect to the analytical model set forth in US Patent Application Publication No. 20160266505.

Alternatively or additionally, several layers can be targeted by performing measurements at an optical plane relative to the target (stage 137) and using measurements of a plurality of pixel positions at the optical plane (stage 138). Concurrently, the application machine learning algorithm 130 is executed to calculate SCOL metrology parameters (stage 135), for example, as in the second variation 100C of the method 100 set forth above. For example, the SCOL metrology parameters may be overlays between N layers, and machine learning algorithms may be applied with respect to the analytical model set forth in US Patent Application Publication No. 20160266505. In one non-limiting example, N=3, the SCOL metrology parameters are overlays between the three layers, and machine learning algorithms can be applied with respect to the analytical model set forth in US Patent Application Publication No. 20160266505. In certain embodiments, any of the measurements may be performed in the field plane (stage 139).

In particular embodiments, method 100 may include deriving a machine learning algorithm (e.g., stages 120 and/or 130) during setup or training (e.g., once) and applying the derived algorithm at runtime (e.g., stage 135) (stage 140), possibly adjusting the derived algorithm (if necessary). Certain embodiments include using simulation, at least in part, to derive a machine learning algorithm (stage 142).

Particular embodiments include a multi-layer metrology target having a plurality of target cells on at least three target layers, each cell having at least one periodic structure in each layer, wherein the periodic structures of each cell are offset relative to each other. Move to the specified offset. The goal is to provide SCOL measurements which are also part of the present invention.

Method 100 provides a number of novel aspects, such as: measuring SCOL targets with more than two overlapping parallel gratings with minimal inaccuracy penalty; combining two-cell target and multi-cell target sampling to accurately measure multiple Overlapping grating targets; targets and measurement methods that comply with all on-die device layout constraints including lateral and vertical constraints without inaccuracy penalties; multi-cell multi-grating targets and measurement methods that Obtain improved throughput and/or effective area; simulation-based target design optimization that considers all process and lithography steps and resulting patterns rather than just two specific desired layers; all-layer patterning based on an analytical model Optimization—The model predicts optimal optical properties that minimize undesired contributions to the signal; uses information across the optical beam (such as reflectivity, differential signal, and/or alignment) to obtain accurate alignment of each layer ; combine the aforementioned per-pixel responses with multi-cell alignment calculations to obtain a calibration against a standard algorithm such that the desired alignment can be obtained from a one or two-cell target; and apply machine learning algorithms to any of the above One. quasi-periodic goals

In the following, examples of quasi-periodic SCOL targets are presented, which are more similar to device patterns and are periodic (but not lattice) on predefined scales. This means that the entire structure cannot be divided into smaller identical structures (unit cells). Note below that in the specific case explained below, the actual device pattern can also be considered a quasi-periodic SCOL target and does not necessarily contain a given shift as shown below. Therefore, the following disclosure enables direct measurement of the overlay of certain device designs even though they are aperiodic.

Machine learning algorithms can be applied to measurements of disclosed targets to enhance the accuracy of result derivation, reduce errors, speed up overall measurement time, reduce the number of cells required, and/or achieve the goal of model-free real-time measurement of similar devices . In certain embodiments, a machine learning algorithm may be trained on a target design based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Figures 5A-5D and 6A-6F are high-level schematic illustrations of a quasi-periodic SCOL target 400 in accordance with certain embodiments of the invention. Target 400 exemplifies a target that satisfies the periodic target requirements of various OVL/alignment technologies (e.g., SCOL, AIM, scanner alignment marks), but is not periodic in the sense that target 400 does not have a unit cell. sexual. Although the target 400 does not have repeating unit cells, the Fourier transform of the target 400 does appear to correspond to certain defined distances corresponding to an effective target spacing on an additional length scale (which can be much larger than the actual fine-scale spacing). Periodicity on length scales. In the analysis of measurements of target 400, random parts that break translational symmetry can be considered "noise" and can be averaged out using measurements over a larger area or using multiple measurements over different target areas. Alternatively or additionally, measurement conditions may be selected to minimize the contribution of translational variation features, or complex signal analysis, target design, and/or hardware may be used to eliminate the contribution of such translational variation characteristics. Individual metrological measurements of target 400 are also considered part of the invention.

In Figure 5A, target 400 is schematically illustrated using a basic pattern of horizontal and vertical lines to indicate the quasi-periodic nature of target 400 ( the X- axis and Y -axis illustrate two measurement axes, the period along the Y- axis can correspond to the standard SCOL target mid-range). Note that in the details, the components 410A to 410F that make up the target 400 each include gaps and cuts. These gaps and cuts modify the components and the target 400 as a whole to avoid being a regular grid, so that the target 400 is based on the grid. , but incorporates irregularities derived from the device pattern, as illustrated in an illustrative manner in Figures 5B and 5C. Figures 6A-6F further elaborate on this aspect by showing object 400 as being composed of blocks 410A-410F, etc. (Figure 6A), which are designed to be schematic representations or abstractions of device patterns (Figure 6B to Figure 6E) and can be combined to form a quasi-periodic target 400 (Figure 6F). It should be emphasized that all block designs are based on a similar periodicity represented by individual grids, but still contain a number of irregularities or aberrations that are symmetrical with respect to the grid, which overall produce Quasi-periodic target 400.

Figure 5B is a high-level schematic illustration of a simplified device layout 420 (eg, a NAND gate layout) that serves as a basis for pattern 410A illustrated in Figure 5C. Pattern 410A includes specified features, for example, pattern 410A can be produced by further simplification such as maintaining columns 411 from device layout 420 and using various types of cutouts 412 similar to the metal lines connecting columns in an actual device layout, and such as Features derived from device layout 420 (alternative pattern to pattern 410E illustrated in Figure 5D). Figure 5A schematically illustrates one of the plurality of patterns represented by 410A, 410B, 410C, 410D, 410E, 410F, etc., which may also be a function of the size of pattern 410A and/or the cutouts 412 in the pattern. Configuration changes. Target 400 may thus be formulated as a superposition of repeating unit cells with different cuts, and/or may be designed to completely lack a defined repeating unit cell. Different cuts in patterns 410A-410F may be selected to represent different logic gates. The illustrated design may be provided for one or more layers of target 400 . Note that the illustrated lines and cuts may be the result of multiple lithography steps (eg, lines may be formed using a pitch multiplication process) and then possibly multiple cut applications. However, the patterning can also be performed in a single lithography step. The illustrated lines and cuts are used to provide a non-limiting example of designated features of the pattern, and may be replaced with features relevant to the device design.

6A schematically illustrates targets 400 as quasi-periodic in the sense that they exhibit periodicity along the Y- axis ( Y- direction periodicity) resulting from the general organization of the wafer, which may be the case in prior art The order of magnitude of the distance between targets) and a regularity along the to 410F etc. can be alternated non-periodically. An evaluation of the degree of irregularity in the design of the target 400 is presented below and is shown to enable the derivation of metrological signals and metrological parameters taking into account the deviations introduced by the irregularity.

One of the important and surprising insights obtained by the inventors from the disclosed analysis is that devices and device segments can also be considered as quasi-periodic targets 400 and therefore considered when considering the effects introduced by their "irregularities" (e.g. regarding strict When periodicity is considered), use the measurement techniques and algorithms presented in this article to directly measure. In addition, the quasi-periodic nature of target 400 enhances the applicability and efficiency of machine learning algorithms for measurement of target 400 .

Figure 6B schematically illustrates schemes 420B, 420C representing schematic layouts of NAND gates and NOT gates, respectively (the background grid is only used to illustrate the periodicity of the pattern and is not an actual part of the pattern). In this exemplary procedure, the M1 pattern is produced using three lithography steps (denoted LELELE, where L represents a lithography step and E represents an etch step, applied to the same physical layer) to give the corresponding M1a, M1b and M1c. Figure 6C only schematically illustrates the M1 pattern common to 420B and 420C. Pattern 410C may be used to represent designs 420B, 420C in object 400. Figure 6D schematically illustrates schemes 420D, 420E representing schematic layouts of OR gates and AND gates respectively and Figure 6E schematically illustrates corresponding M1 patterns 410B, 410D that may be used to represent designs 420D, 420E in target 400. Obviously, additional patterns may be used and incorporated into target 400 according to various performance requirements and optimizations. Note that all of the illustrated designs 410A-410E illustrate the quasi-periodic nature of targets 400 that maintain a large degree of periodicity while introducing irregularities in the pattern corresponding to the particular device design. Figure 6F schematically illustrates a combination of patterns 420B, 420C, 420D that produces a quasi-periodic target 400 (the background grid is only used to illustrate the periodicity of the target and is not an actual part of the target). Note that schemes 420A to 420E serve as one schematic adaptation of a circuit such as that presented by U.S. Patent No. 8,863,063 and they serve as non-limiting examples of possible schemes that can be used to derive corresponding patterns 410A to 410F and other patterns. .

7A and 7B present simulation results of the effect of noise introduced by aperiodic target design on first-order amplitude according to certain embodiments of the present invention. Noise represents irregularities in a substantially periodic structure (which is referred to above as quasi-periodic). The following graphs can be used to estimate the extent to which deviations from periodicity, illustrated in Figures 5A-5D and 6A-6F, degrade the metrology signal derived from target 400. The Fraunhofer approximation is used to calculate the optical plane signal (amplitude) of a grating with irregularities. In OVL measurements, the modification of dS in the first-order noise will roughly change the OVL by dS/A ( A is the measurement sensitivity). Random noise is added to the raster in the form of locations where the amplitude is modified to zero. Random noise may be understood to represent irregularities caused by differences arising from specific patterns 410A to 410F. The calculation is repeated for several different illumination beam positions. The effect of the noise magnitude is calculated on the first diffraction order amplitude distribution as a function of the beam position. In Figure 7A, the error bars indicate the standard deviation of ten different beam positions. All values are normalized relative to the unperturbed intensity. Figure 7B clearly shows the variability between first-order intensities when sampling different locations (corresponding to the error bars of Figure 7A). The different spatial distributions of noise points create an uncertainty of about 0.3% for the amplitude of 2% of the noise magnitude. 7A and 7B illustrate that deviations from strict periodicity result in control noise that represents an irregularity in the target design and can be considered as a factor of inaccuracy in deriving metrology results from target 400. In addition, Figures 7A and 7B provide tools for handling noise due to irregularities in the target structure or in the design of the device used as the target, as suggested below. Machine learning algorithms can be configured on these results and tools to converge quickly and provide accurate derivation.

This inaccuracy can be handled algorithmically (for example, using known symmetry properties of the signal) or by selecting measurement points that counteract symmetry violations. The latter can be accomplished, for example, by automated analysis of zoom masks. Machine learning algorithms can be applied to enhance derivation accuracy, reduce errors, speed up overall measurement time, reduce the number of cells required, and/or achieve model-free real-time metrology measurements for similar devices. In certain embodiments, a machine learning algorithm may be trained on a target design based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Particular embodiments include metrology targets 400 having irregularly repeating units 410A through 410F along at least one direction of the target 400 (possibly two perpendicular directions), wherein the units include one of the features derived from the respective device design. Or multiple sets (different sets) of lines and incisions and other device patterns. For example, element length, line characteristics in the element, and/or cut characteristics in the element may vary along at least one direction of the target 400 . Target 400 may include two or more layers and may provide SCOL measurements, which are also part of this invention.

Figure 8 is a high-level flow diagram illustrating a method 600 in accordance with certain embodiments of the invention. Method 600 may be performed, at least in part, by at least one computer processor. A computer program product and corresponding metrology module are provided, which include a computer-readable storage medium having a computer-readable program embodied thereon and configured to at least partially perform method 600. It also provides target design files and target measurement.

Method 600 may include deriving a plurality of device-like patterns from respective plurality of device designs, wherein the device-like patterns include different sets of lines and cuts as illustrative specified pattern features (stage 615 ), and using the derived device-like patterns as patterns along the target. A measurement target is designed based on the irregular repeating units in at least one direction (stage 620).

Method 600 may include varying at least one of: an element length, line properties in the element, and cut properties in the element along at least one direction of the target (stage 630). At least one direction may include two perpendicular directions of the target. A target may include at least two layers. Method 600 may include possibly using machine learning algorithms to estimate noise resulting from target irregularities that are deviations from strict periodicity and based on specified noise thresholds (stage 632). Design or select an appropriate pattern (stage 634). Method 600 may include estimating a measurement error based on the estimated noise (stage 636), possibly using a machine learning algorithm. Method 600 may include utilizing pattern symmetry properties to estimate and improve the signal received from the pattern (stage 638), as explained above (e.g., by algorithmically, using known symmetry properties of the signal, and/or by selecting cancellations Measurement points of symmetry breaking (e.g., by automatic analysis of reticle reticle) to process the estimated noise). Method 600 may include, for example, after estimation 632, using the reference measurements to learn the overlay signal in order to reduce signal contamination due to target irregularities (stage 639), for example, by using a machine learning algorithm to self-align Signal identification and reduction of signal contamination due to target irregularities.

In certain embodiments, method 600 may include deriving a machine learning algorithm during setup or training (e.g., once) and applying the derived algorithm at runtime (stage 640), possibly adjusting the derived algorithm if necessary ). Certain embodiments include using simulation, at least in part, to derive a machine learning algorithm (stage 642). Avoid offsets in device targets

Returning to the basic SCOL assumption, it is assumed in US Patent Application Publication No. 20160266505 that the measured differential signal (the intensity difference between the first order and the respective negative first order) can be written as D(n)=A(n )·ε(n) , where n is the index of one of the target cells (or target sites) and ε is the lateral offset between two target periodic structures (eg, gratings) in the measurement direction. Since both sensitivity A and relative offset (or OVL) can vary between targets, two parameters should be calculated per target, and thus two measurements with the same sensitivity and OVL are required. If the sensitivity has not changed, OVL(n)=D(n)/A can be used to calculate OVL, but in fact this is not true and a large inaccuracy value is obtained.

To form two informative measurements, a predetermined offset is applied by design. Such offsets can damage the electrical properties of the device and therefore render it unusable in real devices. Since there is only one critical direction for many OVL alignments (as schematically illustrated below in Figures 9 and 10), it can be applied without damaging the device and without affecting the final (after etching) pattern. Offset in the orthogonal direction. In the conventional SCOL algorithm, offset in the orthogonal direction does not help to recover the sensitivity because the device pattern is not symmetrical with respect to a 90° rotation and therefore the measurement of the sensitivity in this direction may not be consistent with the desired Sensitivity related. In the following derivation, a linear approximation to the differential signal is used in a non-limiting way for simplicity. It is expressly stated that all methods disclosed below are also valid for differential signals using higher order approximations and such applications are considered part of the present invention. The following method uses orthogonal offset information to use orthogonal diffraction orders of the same target that do not require any offset and to use an orthogonal target with a different design that has an offset in a non-critical direction (these two options Both enable potentially using actual devices as targets to derive metrology parameters (such as overlays).

Figures 9 and 10 are high-level schematic illustrations of device alignment 97 in accordance with certain embodiments of the invention. For example, FIGS. 9 and 10 may represent an alignment of the contact 711 with the gate 712 . Figures 9 and 10 schematically illustrate that alignment along one direction (critical direction, denoted X ) imposes much more stringent overlay requirements than alignment along a vertical direction (non-critical direction, denoted Y ) (smaller OVL value). Figure 10 also schematically illustrates a first diffraction order signal 98 in the beam plane (for an exemplary central illumination, beam image) of device 97, where along the Y direction (the direction perpendicular to the critical direction) The +1 and -1 diffraction signals are similar to each other (and have rotational symmetry), while the +1 and -1 diffraction signals along the , in intensity).

In the following, rotational symmetry along the non-critical measurement axis is exploited to enable measurements along the critical measurement axis without currently introducing a designed offset along the critical measurement axis. In addition, machine learning algorithms can be applied to measurements to enhance derivation accuracy, reduce errors, speed up overall measurement time, reduce the number of cells required, and/or achieve model-free real-time metrology measurements for similar devices. In certain embodiments, a machine learning algorithm may be trained on a target design based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Figure 11 is a high-level schematic illustration of diffraction orders along a non-critical measurement direction 716 bounding a measurement direction 715, respectively, in accordance with certain embodiments of the present invention. Figure 11 illustrates a simplified model of a grating on grating in one of the two directions (denoted as a grating on grating model 725 and a single grating model 726, since the top grating is aperiodic along the non-critical measurement direction. sexual, see Figure 10).

Note that the simplified model is presented for explanation reasons and does not limit the invention. The gratings 701, 703 will represent any periodic structure, and equivalent models can be used for multi-layer periodic structures. In addition, the measured structure may be a measurement target and/or an actual device. For example, model 725 may be considered to represent a valid two-dimensional periodic structure, while model 726 may be considered to represent a valid one-dimensional periodic structure.

Machine learning algorithms may be configured to improve results relative to any such models (eg, models 725, 726) and may be configured to enhance or replace the use of such models. The results based on the models 725, 726 can be used to guide or train machine learning algorithms and also suggest the applicability and good convergence properties of the machine learning algorithms, and the machine learning algorithms can make it possible to avoid assumptions involved in such models. Certain assumptions.

In model 725, the electric field along the x- direction on the collection beam is the interference between the two diffraction modes illustrated in FIG. Equations are presented and developed to indicate that the two diffraction patterns can be supplemented or replaced by machine learning algorithms. In model 726, the electric field E and the resulting measured intensity I _p are expressed as components 77, and are analyzed analytically in US Patent Application Publication No. 20160266505, which are generated after passing through the middle layer 702 and the upper layer 701. Reflecting the lower grating 703, which upper layer contains the upper grating along non-critical (non-gauge) directions and therefore lacks periodicity, the model 726 can be supplemented or replaced by a machine learning algorithm.

As mentioned above, more complex models and calibration functions can be implemented using the same approach and are considered part of the present invention. Note that the orthogonal diffraction order can also be used to calculate the geometric properties of the target (for example: critical dimensions) with or without optical modeling. Machine learning algorithms may be configured to improve results relative to any such models and may be configured to enhance or replace the use of such models. Results based on such models may be used to guide or train machine learning algorithms and may avoid some of the assumptions involved in such models.

The disclosed analysis and machine learning methods (see also method 800 below) may be implemented in various ways to derive metrological measurements from various devices and target designs (an overlay of such metrological measurements is presented as a non-limiting example). As a non-limiting example, Figure 12 schematically illustrates an exemplary configuration for application of the method.

Figure 12 is a high-level schematic illustration of a target 700 incorporated into an offset-free device portion in accordance with certain embodiments of the present invention. The target 700 can be designed to eliminate the need for pixels along the critical OVL dimension 715 (in-cell 710) by using additional cells 720 with patterns and offsets in a different direction 716 (e.g., perpendicular to the critical direction of the cell 710). Provides a sensitivity calculation in the case of offsets in element 710). Figure 12 illustrates a three-cell design in two layers in a non-limiting manner, but can be extended to a multi-layer design (as explained above in the present invention), to quasi-periodic targets and device (as explained above). Objective 700 achieves overlay calculations along critical direction 715 without introducing a given offset in this direction. Note specifically that cell 710 may be understood to represent at least a portion of an actual device design, and the disclosed methods thus enable the introduction of offsets into the device design at least along critical directions of the device, possibly without introducing into the device design. Measure the device directly without introducing any offset.

U.S. Patent Application Publication No. 20160266505 presents an analysis method suitable for measuring the center cell 710 and measuring the other (eg, two) cells 720 with a given offset ± f ₀ for Y differential signal calculation, thereby Provides equations expressing the sensitivity A of a target approximated by a function of a second nearby target. Machine learning algorithms can be used to enhance or replace these calculations, possibly achieving any of the following: reduction in the number of cells per target, extracting more information and/or extracting information faster from self-measurement, and about the machine Improved objective design of parameters for learning algorithms. Machine learning algorithms can be trained on target designs based on metrological simulations to match a behavior of the target design to the behavior of a specified device pattern.

Figure 13 presents a table of exemplary simulation results with resulting sensitivity values for different combinations of first cell design 710 and second cell design 720, respectively, in accordance with certain embodiments of the present invention. The table presents sensitivities as inaccuracy values (in nm, based on simulations), with high sensitivity being above approximately 1 nm and low sensitivity being below approximately 1 nm, and demonstrates the effectiveness of the disclosed method. Test methods 800 for fifteen different target designs and ninety program variations. The presented calculations described in U.S. Patent Application Publication No. 20160266505 ensure the applicability and good convergence properties of the machine learning algorithm, while allowing the machine learning algorithm to improve the results and possibly avoid assumptions involved in the analysis method. Certain assumptions.

The number of cells in a SCOL target can be reduced using the methods described above. For example, N+1 cells can be used to measure the relative shifts of N features in the same layer instead of 2N cells - two cells are used to calculate the first feature sensitivity and OVL, and all Other designs may have a single cell for which a calibrated sensitivity function may be used relative to the first design. In certain embodiments, machine learning algorithms may be used to further reduce the number of cells, possibly even to a single cell per target, where the machine learning algorithm is configured to achieve model-free real-time optical alignment of single cell targets. Measurement. For example, after calculating the calibration function based on simulation, measurement and/or machine learning algorithms, the metrology procedure can also be improved. The running sequence can include real-time measurement of individual cells and use of orthogonal direction signals to calibrate using the calculated calibration function. sensitivity. Note that in the case of unstable procedures, several different calibration targets can be used.

Figure 14 is a high-level flow diagram illustrating a method 800 for measuring overlay without introducing a defined shift along a critical direction, in accordance with certain embodiments of the present invention. Method 800 may be implemented, at least in part, using machine learning algorithms. Method 800 may be implemented, at least in part, by at least one computer processor (eg, in a metering module). Particular embodiments include computer program products that include a computer-readable storage medium having computer-readable programs embodied thereon and configured to perform relevant stages of method 800 . Certain embodiments include target design files for respective targets designed by embodiments of method 800.

Method 800 includes simultaneous measurement overlap while avoiding prior art techniques that introduce a predetermined offset along a critical measurement direction into at least one of the target cells (stage 805). Measurements can be performed in a model-free manner, possibly using machine learning algorithms. Method 800 may include using the intensity of the diffraction order orthogonal to the critical measurement direction and applying a machine learning algorithm to calibrate the sensitivity parameters using offsets in an orthogonal non-critical measurement direction (stage 810). Alternatively or in addition, method 800 may include designing a reference calibration target on the reticle (stage 820) and applying a machine learning algorithm to calibrate the sensitivity parameters using offsets in the calibration target (stage 825). Method 800 may include selecting parameters of the reference target based on the model to reduce inaccuracy (stage 830). Method 800 may include using at least one additional target cell instead of at least one target cell to measure sensitivity, for example, by introducing an offset (along either or both critical and non-critical directions) (Stage 815). The at least one additional target cell may be adjacent to the target cell and/or configured as a separate calibration target.

Method 800 may include designing metrology targets that are incorporated into at least a portion of a device design where the cells have offset in a non-critical direction and the device portion exhibits no offset (stage 840). In certain embodiments, multiple portions of an actual device may be used and combined into a single metrology measurement. Method 800 further includes selecting multiple device design portions to generate a derived image from respective optical images of those portions. A light plane image, the derived light plane image satisfying a specified criterion (eg, with respect to periodicity and/or estimated noise) (stage 845). For example, one optical beam image used in the OVL calculation may be an average of several optical beam images measured at different, possibly different, device regions 50A. The selection of combinations may be predefined or performed automatically to effectively select signals that provide specific characteristics (eg, most closely resemble a signal derived from a periodic target, exhibit lowest noise levels, etc.). The method 800 may include using a machine learning algorithm to optimize the target design based on approximation assumptions (stage 850 ), for example by using the machine learning algorithm to reduce variation in the form factor of the periodic structure (stage 855 ). Offsets may be introduced along orthogonal non-critical measurement directions of the device design and/or in adjacent non-device additional cells and/or in calibration cells (eg, scribe lines).

The following aspects are provided by method 800 and the disclosure above. OVL sensitivity calibration may be performed based on real-time information from additional diffraction orders (which may include orthogonal diffraction orders and may be derived by machine learning algorithms). OVL sensitivity calibration may be performed based on real-time information from a second target with a different design and/or from additional targets with a different design and/or diffraction order reflectivity. These enable the use of OVL targets with no offset in critical measurement directions in accordance with the disclosed measurement methods (enabling direct device measurements without degradation of electrical functionality). Examples of targets presented above include any of the following: a single-cell SCOL target for one direction (x or y) where sensitivity is calculated based on orthogonal direction reflectance; a triple-cell SCOL target SCOL target, which consists of two standard cells for a first direction and a third cell for an orthogonal direction; contains N+1 cells (instead of 2N cells) along a single direction One SCOL objective of N designs—a first design with two cells and all other designs with a single cell per design; and a design with N+2 cells (instead of 2N ) cells along a critical One of N Designs in Direction SCOL Objective—One first design has two cells in the orthogonal non-critical direction and all other designs have a single cell per design (without offset). It is important to note that all cells must not be adjacent to each other, for example the target may be a combination of cells located within the active area of the device and cells located in the periphery of the active area of the device.

The present invention further provides OVL-free model target measurement using at least one of the following: no established offset; no defined unit cells; multiple (two or more) overlapping patterning (i.e., possible applications Different lithography steps for the same physical layer); and using a SCOL-like algorithm (run-time model-free method) for device pattern measurement. The method enables optical measurement of the device pattern after photoresist development as well as optical measurement of the final and post-etched device pattern. The present invention further provides targeted design optimization based on metrological simulation to match specific device pattern behavior and metrological simulation combined with lithography and/or procedural simulation to achieve targeted design optimization to match specific device pattern behavior. Finally, model-free real-time optical OVL measurements using a single cell and a combined OVL and OCD (optical critical dimension) target using a model-free OVL algorithm are provided (multiple hardware configurations may be required for measurement, as above) explained in the text).

Figure 15 is a high-level schematic illustration of a composite device object 700 in accordance with certain embodiments of the invention. Figure 15 illustrates in a schematic manner the combination of the concepts disclosed above to produce direct metrology measurements in a device area 50 (see method 900 below). As illustrated in Figure 15, a device region may be considered to be multi-layered, quasi-periodic and have no offset and/or have offset only in non-critical directions. This understanding is surprising because metrology target design is often very different from device design. However, the inventors have discovered that from this perspective and/or by selecting specific device areas based on these criteria, the actual device area can be successfully processed and measured as a metrology target or a portion thereof and provide a direct link to the device. Available measurement results related to characteristics. By considering and/or selecting the device area to be multi-layered, quasi-periodic (in the sense explained above) and without offsets and/or with offsets only in non-critical directions - it can be achieved by respectively Application methods 100, 600, and 800 (possibly enhanced using machine learning algorithms 150) use respective device regions as targets or target portions, as schematically illustrated in Figure 15.

For example, the target 700 may include at least one region 50A of the device 50 as a portion 710 of the measured target in which offsets are not introduced (at least along critical directions with respect to the functionality of the device) and have a Sensitivity parameters and/or calibration functions are used to derive adjacent cells 720 at a given offset for device overlay. 15 schematically illustrates selection of one region 50A as an option for a target portion 710 adjacent to a cell 720, and selection of a plurality of regions 50A as an option for a target portion that may but not necessarily be adjacent to a cell 720. Region 50A may be selected to generate a derived beam plane image (e.g., an average or a weighted average of the respective beam images) from the respective beam images of region 50A that is sufficient for the purpose. A specified criterion, such as a noise threshold, a periodicity threshold, or any algorithmic threshold, is optimized in view of the quality of the derived signal and the associated overlay derivation.

Alternatively or in addition, the target 700 may include at least one region 50B of the device 50 as the target 700 or a portion of the target having a predetermined offset introduced at the particular region 50B along a non-critical direction of the device design. These defined offsets may be selected to provide useful metrology information (eg, sensitivity parameter A ) without compromising device performance (see explanation and derivation above corresponding to Figures 9-11). Target 700 may further include neighboring cells 720 having established offsets that may be used to derive device overlay based on precomputed sensitivity parameters and/or calibration functions.

Direct device measurements may further utilize calibration targets 750 set, for example, on the reticles, which calibrate any of the effects of multilayering, quasi-periodicity, and sensitivity, as explained above. Additionally, methods 100, 600, and/or 800 (perhaps at least partially performed and/or enhanced using machine learning algorithms 150) may be collaboratively implemented as one of the methods 900 set forth below, such that methods 100, 600, and/or 800 can be implemented without introducing along Direct metrological parameter measurement of multi-layered and non-periodic devices is carried out when the critical direction of the device design is shifted.

Figure 16 is a high-level flow diagram illustrating an integrated method 900 for measuring device overlay directly on a device, in accordance with certain embodiments of the present invention. Method 900 may be performed and/or enhanced, at least in part, using machine learning algorithms. Method 900 may be implemented, at least in part, by at least one computer processor (eg, in a metering module). Particular embodiments include computer program products that include a computer-readable storage medium having computer-readable programs embodied thereon and configured to perform relevant stages of method 900 . Particular embodiments include target design files for respective targets designed by embodiments of method 900.

Method 900 may include using reference calibration targets and/or device adjacent cells with established offsets to enable direct measurements of device portions without introducing offsets into the device design (stage 910), e.g., implementing Method 800. Method 900 may include applying a machine learning algorithm to calibrate sensitivity using at least one of: introducing an offset along a non-critical direction, using adjacent target cells with an introduced offset, and using scribing. Online sensitivity calibration target (stage 915), as explained above.

Method 900 may include extending the cell design to multi-layer measurements (stage 920), for example, implementing method 100 in any of variations 100A to 100D of method 100 (possibly enhanced by a machine learning algorithm). Method 900 may include configuring additional targets to provide layer-specific metrology parameters using multiple layers ( N ) of target cells (with N > 2 overlapping layers) including at least one of the following: N pairs of cells, each pair having opposite offsets at a different layer; N cells having a chosen offset; N-1 or less cells having an offset configured to utilize light A selected offset of the information; and a calibration target located next to the overlay target, which has as low as 2 cells (stage 925), as explained above. In certain embodiments, the targets may include a single cell per target, wherein the machine learning algorithm is further configured to achieve model-free real-time optical overlay measurements of the single cells.

Method 900 may include directly measuring periodic design patterns while managing and limiting resulting inaccuracies (stage 930), for example, implementing method 600. Method 900 may include measuring metrology parameters from at least a portion of a device design, the at least portion being selected to have a plurality of irregular repeating units along at least one direction of the portion, the plurality of irregular repeating units having different sets of lines, and The cutout serves as an exemplary designated feature (stage 935), as explained above.

Method 900 may include integrating derivation of offset-free, multi-layer and quasi-periodic measurement algorithms (stage 940) to measure metrology parameters directly on the device (stage 950).

Method 900 may further include selecting parameters adjacent to the target cell and/or sensitivity calibration target to reduce inaccuracy according to a model of the inaccuracy (stage 955), as illustrated above.

Corresponding metrology targets include at least a portion 710 of a device design selected to have a plurality of irregular repeating units along at least one direction of the portion (eg, as schematically illustrated in patterns 420A through 420E), the A plurality of irregular repeating units having different sets of lines and cuts (e.g., as schematically illustrated in object 400); and a plurality of additional cells including multiple layers of alignment cells (e.g., as schematically illustrated in objects 200 and 300) exemplified) and sensitivity calibration cells (eg, as schematically exemplified in targets 700 and 750). Multiple layers of calibration cells may include any of the following (see Figure 1A): N pairs of cells, each pair having opposite biases at a different layer; N cells having selected biases shift; N cells with selected offsets configured to exploit the optical information (or possibly fewer cells, depending on the measurement conditions and algorithm used); and located next to the overlay target The N-cell calibration target has N-1 cells and can be as low as 2 cells depending on the calibration conditions and algorithm complexity. The sensitivity calibration cells may include at least two additional cells with a predetermined offset along the critical measurement direction of the at least one target cell, e.g., with an orthogonal critical measurement direction relative to the at least one target cell. extra cells. At least two additional cells may be adjacent to the device portion as additional cells 720 . Individual metrological measurements of the disclosed objects are also considered part of this invention.

Figure 17 is a high-level flow diagram illustrating a method 960 of applying a machine learning algorithm to any of the disclosed methods in accordance with certain embodiments of the invention. Method 960 and/or its stages may be integrated into any of the methods disclosed above, and may be performed and/or enhanced, at least in part, using machine learning algorithms. Method 960 may be implemented, at least in part, by at least one computer processor (eg, in a metering module). Particular embodiments include computer program products that include a computer-readable storage medium having computer-readable programs embodied thereon and configured to perform the relevant stages of method 960 . Certain embodiments include target design files for respective targets designed by embodiments of method 960.

Method 960 and/or any of methods 100, 600, 800, and 900 may include any of the following stages: applying a machine learning algorithm to calibrate measurement sensitivity (stage 962); using a machine learning algorithm to Reduce the number of cells required (stage 965); train the machine learning algorithm on the simulated target (stage 970); use the training results to optimize the target design to match the specific device pattern (stage 975); compare the behavior of the target design with Behavioral matching of specific device patterns (stage 977); and use of machine learning algorithms to achieve model-free real-time measurements of individual cells (stage 980).

Aspects of the present invention are described above with reference to flowchart illustrations and/or partial drawings of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each portion of the flowchart illustrations and/or portions of the drawings, and combinations of portions of the flowchart illustrations and/or portions of the drawings, can be implemented by computer program instructions. Such computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing device to produce a machine that is executed by the processor of the computer or other programmable data processing device. The instructions create means for implementing the functions/actions specified in the flowchart and/or one or more partial diagram sections.

These computer program instructions can also be stored in a computer-readable medium that can direct a computer, other programmable data processing equipment, or other devices to function in a specific manner such that the computer-readable medium stored in the computer-readable medium The instructions in the media generate a fabricated object containing instructions to implement the functions/actions specified in the flowchart and/or one or more partial schematic portions.

The computer program instructions may also be loaded onto a computer, other programmable data processing equipment, or other device to cause a series of operating steps to be performed on the computer, other programmable equipment, or other device to produce a A computer-implemented program such that instructions that execute on a computer or other programmable device provide a program for performing the functions/actions specified in the flowchart and/or one or more partial diagram portions.

The foregoing flowcharts and figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the invention. In this regard, each portion of the flowchart or partial diagrams may represent a module, segment, or portion of code that includes one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions described in this section may occur out of the order depicted in the figures. For example, in fact, two parts of a consecutive presentation may be performed substantially simultaneously, or the parts may sometimes be performed in the reverse order, depending on the functionality involved. It should also be noted that each portion of the drawings and/or flowchart illustrations, and combinations of portions of the drawings and/or flowchart illustrations, may be configured by application-specific hardware-based systems or special hardware that perform the specified functions or actions. Implemented using a combination of hardware and computer instructions.

In the above description, an embodiment is an example or implementation of the present invention. The various appearances of "one embodiment," "an embodiment," "a particular embodiment," or "certain embodiments" are not necessarily all referring to the same embodiment. Although various features of the invention may be described in the context of a single embodiment, they may also be provided individually or in any suitable combination. Rather, although the invention is described herein for clarity in the context of separate embodiments, the invention can also be practiced in a single embodiment. Particular embodiments of the invention may contain features from different embodiments disclosed above, and certain embodiments may incorporate elements from other embodiments disclosed above. The disclosure of elements of the invention in the context of a particular embodiment should not be construed as limiting the use of such elements alone in that particular embodiment. Furthermore, it is to be understood that the invention may be embodied or practiced in various ways and that the invention may be practiced in specific embodiments other than those outlined in the description above.

The present invention is not limited to the drawings or corresponding descriptions. For example, the process need not move through each illustrated box or state or in the exact same order as illustrated and described. Unless otherwise defined, technical and scientific terms used herein have the meaning generally understood by one of ordinary skill in the art to which this invention belongs. Although the present invention has been described with respect to a limited number of embodiments, these embodiments should not be construed as limiting the scope of the invention but as illustrating some of the best embodiments. Other possible variations, modifications and applications are also within the scope of the invention. Accordingly, the scope of the invention should not be limited by what has so far been set forth, but rather by the scope of the appended patent applications and their legal equivalents.

50: Device area/Device 50A: Device area/Area 50B: Area 60: Wafer 77: Component 90: Overlay target and algorithm/Algorithm 97: Device alignment/Device 98: First diffraction order signal 100: Method/Measurement method/Measurement method 100A: Method/Modified scattering measurement overlay algorithm/Analysis method/Variation of the method 100B: Method/First method/Procedure/Differential signal analysis method/Analysis method/Modification of the method First variation/variation of method 100C: Method/Second method/Procedure/Analysis method/Second variation of method/Variation of method 100D: Method/Analysis method/Variation of method 102: Stage 105: Stage 110: Stage 115: Stage 120: Stage 130: Stage 132: Stage 135: Stage 137: Stage 138: Stage 139: Stage 140: Stage 142: Stage 150: Machine learning algorithm 155: Reduce the number of cells required 200: Calibration Target / target / multi-layer target / multi-cell target 201: target 205: signal / differential signal 210: layer / cell layer 220: cell / target cell 235: overlay 290: overlay target / target 300: cell Element/Target/Multi-layer target/Three-layer target/Two-cell element target 310: Top layer/layer/Top grating 320: Middle layer/layer/Middle grating 330: Bottom layer/layer/Bottom grating 400: Quasi-periodic scattering measurement Overlay target/target/quasi-periodic target/measurement target 410: component 410A: component/block/pattern/design/irregular repeating unit 410B: component/block/pattern/design/irregular repeating unit 410C: component/ block/pattern/design/irregular repeating unit 410D: component/block/pattern/design/irregular repeating unit 410E: component/block/pattern/design/irregular repeating unit 410F: component/block/pattern/ Design/irregular repeating unit 411: column 412: cutout 420A: scheme/pattern 420B: scheme/design/pattern 420C: scheme/design/pattern 420D: scheme/design/pattern 420E: scheme/design/pattern 600: method 615: Stage 620: Stage 630: Stage 632: Stage/Estimation 634: Stage 636: Stage 638: Stage 639: Stage 640: Stage 642: Stage 700: Target/Composite Device Target 701: Grating/Upper Layer 702: Middle Layer 703: Grating /lower grating 710:cell/center cell/first cell design/part/target part 711:wafer/component 712:gate 715:critical overlay size/critical direction 716:direction 720:cell/second Cell Design 725: Grating over Grating Model/Model 726: Single Grating Model/Model 750: Calibration Target/Target 800: Method/Procedure Change Test Method 805: Stage 810: Stage 815: Stage 820: Stage 825: Stage 830: Stage 840: Stage 845: Stage 850: Stage 855: Stage 900: Method/Integrated Method 910: Stage 915: Stage 920: Stage 925: Stage 930: Stage 935: Stage 940: Stage 950: Stage 955: Stage 960: Method 962 :Phase 965:Phase 970:Phase 975:Phase 977:Phase 980:Phase E: Electric field :First-order signal :field :Field I:Illumination M1:Pattern﹢f ₀ :Predetermined offset/predefined offset﹣f ₀ :Predetermined offset/predefined offset

For a better understanding of embodiments of the invention and to show how embodiments of the invention may be practiced, reference will now be made, by way of example only, to the accompanying drawings, in which like numerals designate corresponding elements or sections throughout.

In the attached picture:

Figure 1A is a high-level schematic overview illustration of a multi-layered target and a method of measuring the same, in accordance with certain embodiments of the present invention.

FIG. 1B is a high-level schematic illustration of two types of multi-layer targets and measurement methods according to certain embodiments of the invention.

Figure 2 is a high-level schematic illustration of a multi-layered object in accordance with certain embodiments of the present invention.

Figures 3A and 3B are high-level schematic illustrations of multi-layered objects in accordance with certain embodiments of the invention.

Figure 4 is a high-level flow diagram illustrating a method according to certain embodiments of the invention.

Figures 5A-5D and 6A-6F are high-level schematic illustrations of quasi-periodic SCOL targets in accordance with certain embodiments of the invention.

7A and 7B present simulation results of the effect of noise introduced by aperiodic target design on first-order amplitude according to certain embodiments of the present invention.

Figure 8 is a high-level flow diagram illustrating a method according to certain embodiments of the invention.

Figures 9 and 10 are high-level schematic illustrations of device alignment according to certain embodiments of the invention.

Figure 11 is a high-level schematic illustration of leading diffraction orders along non-critical and critical measurement directions in accordance with certain embodiments of the invention.

Figure 12 is a high-level schematic illustration of an objective of incorporating an offset-free device portion in accordance with certain embodiments of the present invention.

Figure 13 presents a table of exemplary simulation results of resulting sensitivity values with different combinations of first and second cell designs, in accordance with certain embodiments of the present invention.

Figure 14 is a high-level flow diagram illustrating a method of measuring overlay without introducing a defined shift along a critical direction, in accordance with certain embodiments of the present invention.

Figure 15 is a high-level schematic illustration of a composite device object in accordance with certain embodiments of the invention.

Figure 16 is a high-level flow diagram illustrating an integrated method for measuring device overlay directly on the device, in accordance with certain embodiments of the present invention.

Figure 17 is a high-level flow diagram illustrating a method of applying a machine learning algorithm to any of the disclosed methods in accordance with certain embodiments of the invention.

100:Method/Measurement method/Measurement method

100A: Methods/modified radiometric overlay algorithms/analysis methods/variations of methods

100B: Method/first method/program/differential signal analysis method/analysis method/first variation of the method/variation of the method

100C: Method/Second method/Procedure/Analytical method/Second variation of method/Variation of method

100D: Methods/Analytical methods/Variations of methods

102: Stage

105: Stage

110: Stage

115: Stage

120: Stage

130: Stage

132: Stage

135: Stage

137: Stage

138: Stage

139: Stage

140: Stage

142: Stage

Claims (43)

A metrology target comprising: at least one portion of a device design having N > 2 overlapping layers, the at least one portion being selected to have a plurality of irregular repeating units along at least one direction of the portion, the plurality of Irregular repeating units having specified characteristics, and a plurality of additional cells including at least multiple layers of cells and sensitivity calibration cells, wherein the multiple layers of cells include at least one of the following: N cells, which having a selected offset; N or less cells having a selected offset configured to utilize pupil information; and N cell calibration targets located between N-1 and Next to an overlay target between two, the cells are configured according to parameters of at least one machine learning algorithm applied to the measurement and/or simulation of the measurement target. The metrology target of claim 1, wherein the sensitivity calibration cells comprise at least two target cells with introduced offsets adjacent to the device portion. The measurement objective of claim 2, wherein the introduced offsets are orthogonal to a critical direction of the device portion, and wherein the device portion does not have an established offset along the critical direction of the device portion. The measurement target of claim 2, wherein the parameters of at least one of the adjacent target cells and the sensitivity calibration target are selected based on the parameters of the at least one machine learning algorithm, based on Reduce the inaccuracy according to a model of the inaccuracy. Such as the measurement target of claim 1, wherein the sensitivity calibration cells are located on the scribing line. The measurement objective of claim 1, wherein the at least one portion includes a plurality of device design portions selected to generate a derived pupil plane image from respective pupil images of the portions, the derived pupil plane image The pupil plane image satisfies a specified criterion. Such as the measurement target of claim 1, wherein the specified features include at least one set of lines and cuts. A target design file with a measurement target as in any one of claims 1 to 7. A measurement of a measurement objective such as any one of claims 1 to 7. A metrology method, which includes: configuring a multi-layer metrology target to have a plurality of M target cells located on at least three target layers, N

M, each cell has at least one periodic structure in each layer, and the periodic structures of each cell are configured to be offset relative to each other by a specified offset, and the scattering amount is measured from the multi-layer metrology target The measurement method measures at least M differential signals, and applies at least one machine learning algorithm to the differential signals and the specified offsets, by comparing the SCOL measurement parameters with the differential signals and the The SCOL metrology parameters are calculated from the M measurements of the multi-layer metrology target by solving a set of M equations related to the specified offset. The method of claim 10, wherein the multi-layer metrology target includes M<2N cells, and wherein the application of at least one machine learning algorithm is further configured to extract overlay information from the M<2N cells. The method of claim 10, further comprising training the at least one machine learning algorithm on the metrology simulation-based target design of the multi-layer metrology target to match a behavior of the target design with a specified device pattern behavior. The method of claim 10, wherein the at least one machine learning algorithm is derived during setup and/or training and applied at runtime. The method of claim 10, wherein the multi-layer metrology target includes a single cell per target, and wherein the at least one machine learning algorithm is further configured to achieve model-free real-time optical overlay measurement of the single cell. The method of claim 10, wherein the SCOL measurement parameters are overlays between the N layers. The method of claim 10, wherein the application of the at least one machine learning algorithm is executed sequentially for consecutive layers to calculate the SCOL measurement parameters. The method of claim 10, wherein the measurements are performed simultaneously for the layers by performing the measurement at a pupil plane relative to the target and using measurements of a plurality of pixel positions at the pupil plane. Apply the at least one machine learning algorithm to calculate the SCOL measurement parameters. The method of any one of claims 10 to 17, which is at least partially executed by at least one computer processor. A computer program product comprising a computer-readable storage medium having a computer-readable program embodied thereon and configured to at least partially perform the method of any one of claims 10 to 17. A measurement module including the computer program product of claim 19. A target design file for a target designed according to the method of any one of claims 10 to 18. A metrological measurement of an object designed according to the method of any one of claims 10 to 17. A multi-layer metrology target comprising a plurality of target cells located on at least three target layers, each cell having at least one periodic structure in each layer, wherein the periodic structures of each cell are relative to each other The offset is up to a specified offset, wherein the cells are configured according to parameters of at least one machine learning algorithm applied to the measurement and/or simulation of the measurement target. A scattering measurement overlay (SCOL) measurement of a multi-layer measurement target as claimed in claim 23. A metrology method that includes calibrating at least one sensitivity parameter without moving along a critical measurement direction to at least one target by applying at least one machine learning algorithm to calibrate at least one sensitivity parameter using an offset in at least one of: Measuring at least one measurement parameter in the at least one target cell with a predetermined offset introduced in the cell: (i) an orthogonal non-critical measurement direction; and (ii) dividing the at least one target cell at least one additional target cell outside. The method of claim 25, further comprising training the at least one machine learning algorithm on the target design based on metrological simulation to match a behavior of the target design with a specified device pattern behavior. The method of claim 25, wherein the offsets in the orthogonal direction are introduced into at least one additional target cell in addition to the at least one target cell. The method of claim 27, wherein the at least one additional target cell is adjacent to the at least one target cell. The method of claim 27, wherein the at least one additional target cell is positioned on a calibration target on the score line. The method of claim 27, further comprising selecting parameters of the at least one additional target cell based on parameters of the at least one machine learning algorithm to reduce the inaccuracy according to a model of the inaccuracy. The method of claim 25, wherein the at least one target cell includes at least a portion of a device design. The method of claim 31, further comprising introducing the offsets in the orthogonal direction into at least one additional target cell adjacent to the at least one target cell according to parameters of the at least one machine learning algorithm. The method of claim 31, wherein the offsets are introduced along the orthogonal non-critical measurement direction of the device design according to parameters of the at least one machine learning algorithm. The method of claim 31, further comprising introducing the offsets in at least one calibration target positioned on the reticle. A computer program product comprising a computer-readable storage medium having a computer-readable program embodied thereon and configured to at least partially perform the method of any one of claims 25 to 34. A metrology target comprising: at least one target cell that does not have a predetermined offset along a critical measurement direction of the at least one target cell, and at least two additional cells that have a predetermined offset along the at least one critical measurement direction. a given offset of the critical measurement direction of a target cell, The predetermined offsets are derived based on parameters of at least one machine learning algorithm applied to the measurement and/or measurement simulation of the measurement target. The measurement target of claim 36, wherein the at least two additional cells have an orthogonal critical measurement direction relative to the at least one target cell. The measurement target of claim 36, wherein the at least two additional cells are adjacent to the at least one target cell. The measurement target of claim 36, wherein the at least two additional cells are calibration targets located on the score line. The measurement target of any one of claims 36 to 39, wherein the at least one target cell includes at least a portion of a device design. A metrological measurement of an object as in any one of claims 36 to 40. A metrology target, which includes at least one target cell, the at least one target cell does not have a predetermined offset along a critical measurement direction of the at least one target cell and has a predetermined offset along the at least one target cell. A predetermined offset in a non-critical measurement direction, wherein the predetermined offset is derived based on the parameters of at least one machine learning algorithm applied to the measurement and/or measurement simulation of the measurement target. A metrological measurement of an object as claimed in claim 42.